Comparing CORINE Land Cover with a more detailed database in Arezzo (Italy). Javier Gallego JRC, I-21020 Ispra (Varese) ITALY e-mail: javier.gallego@jrc.it Keywords: land cover, accuracy assessment, area estimation Abstract The Italian Statistical Institute (ISTAT) promoted a pilot project for a land use / land cover database at 1:25.000 scale over a test area in the Province of Arezzo, 2000 km 2 wide. The nomenclature was derived from Corine Land Cover with two additional levels. In this paper we analyse the impact of the difference in scale of maps and minimum size of mapping units between CORINE Land Cover and the ISTAT land cover map. A straightforward computation of confusion matrices gives a pessimistic view of the agreement between both land cover maps, but most of the apparent disagreement is due to different scales. An assessment with a simplified nomenclature shows that there is a very good agreement for main classes with two exceptions: pasture and non-forest natural vegetation, that represent only a limited area (less that 1% for pasture). The disagreement for these two categories is partly due do the difficulties in interpreting the nomenclature. 1 Introduction When CORINE Land Cover (CLC) is compared with land cover information at a different scale, premature conclusions could be drawn if a straightforward comparison is seen as a quality assessment. If the comparison between two geospatial land cover data layers takes into account their different scales, the conclusions change. We give a procedure to perform such a comparison trying to remove the effects of different scales and collocation inaccuracy. We also illustrate the fact that CLC, should not be directly used for area estimation by direct polygon area measurement. However CLC, used as covariable, is extremely useful for land cover area estimation (Gallego et al, 1999) 1.1 The Arezzo ILC database and CORINE Land Cover ISTAT, the National Statistical Institute of Italy, has developed a land use-land cover database on an area of 200,000 hectares in the Arezzo province, in central Italy, with the geometric accuracy of 1:25,000 scale. Below we call this land cover map ILC. The main purpose of this database was testing the possible extension at country level (ISTAT, 1998). The pilot project was partially funded by the European Union (EU). The territory is divided into polygons classified into land use - land cover classes. The nomenclature has five levels; the three first levels coincide with the CORINE land-cover nomenclature (EC, 1993); a fourth level is added for urban areas and a forth and a fifth level are added for forests and semi-natural areas.
The project used digital black and white ortho-photos with a resolution of one meter and three Landsat-TM quarter-scenes, taken on the following dates: 10/05/97, 26/10/97, 06/05/98. The photo-interpretation was carried out drawing polygons directly on screen with data at the level of census sections as ancillary information. The minimum mapping unit was 1,56 hectares (1 hectare for urbanised areas). The polygon layer was integrated with a vector layer composed of linear and point features, referring to railways, highways, major roads and rivers and topographic names. Linear and point features were digitised on screen taking into account the digital ortho-images geometry, after identification on 1:250.000 scale maps. The database has undergone several quality checks, including independent photo-interpretation of samples of arcs, polygons and points (Napolitano et al, 2000). Figure 1 gives a visual overview of both land cover maps (ILC and CLC) and Table 1 shows their main differences. In the pilot area there is a dominant presence of forest with significant agricultural and urban areas in the valleys. Urban Arable Perm crops Pasture Heterogeneous Forest Nat. Veg. Open spaces Water CORINE Land Cover ISTAT Land Cover Figure 1: CORINE and ILC land cover maps recoded into 9 classes CORINE Land Cover Arezzo ILC minimum mapping unit 25 ha 1.56 ha (1 ha for urban areas) location accuracy 100 m 25 m Area covered >3 Million km 2 2000 km 2. Extension to 300,000 km 2 proposed. Nomenclature 3 levels, 44 classes same with 2 additional levels. Table 1: Main differences between CLC and ILC
2 Comparing total area per class Land cover maps are sometimes used for area estimation of a land cover class by simply adding the area of the polygons labelled as belonging to that class. This approach is rather naïve and can lead to a serious bias if the mapping scale is not detailed enough or the thematic accuracy is not very high (Gallego et al, 1999). We can compare the total area obtained by this method from CORINE Land Cover and the ILC map for some groups of land cover classes (Table 2). On the other hand interpreting the strong differences that appear for some classes as a sign of poor accuracy of CLC would be also naïf. These differences appear mainly because CLC respects its scale specifications. For example urban areas under 25 ha do not appear in CLC but may appear in ILC and certainly should be counted for statistics. Interpreting the total area of ILC as statistics is also inadequate. Isolated buildings under 1 ha should not appear in ILC and should be counted as artificial for land cover statistics. ILC CORINE Artificial 84.9 57.1 Arable land 375.9 234.8 Permanent crops 141.0 99.7 Pastures 21.8 66.9 Heterogeneous agriculture 110.8 328.2 Forest 1117.6 1029.5 Natural vegetation 132.4 167.2 Open spaces and wetland 1.7 5.5 Water 5.3 2.3 Table 2: Total area in km 2 per land cover class from CORINE and ILC land cover maps Part of the disagreement can be attributed to the fact that the images were taken on different dates, part can be due to the scale effect and part to different thematic accuracy levels. Compared with CORINE Land Cover, the area labelled as heterogeneous agriculture is significantly reduced and attributed to pure classes. This would explain, at least partly, the increase of arable land, permanent crops and forest. However, if we remove the polygons coded in CORINE Land Cover as heterogeneous, we still have 26.000 ha classified as arable land by ILC, i.e. 11% more than CORINE Land Cover. We shall see below that the thematic disagreement for arable land is very low; therefore a significant part of the difference in polygon area comes from the scale effect. Another part of the disagreement may be due to a different interpretation of the nomenclature or to photo-interpretation errors, but it is difficult to know the impact of each source of disagreement without a suitable ground survey. We try below to resolve the impact of each source of disagreement.
3 Statistical analysis of the differences between the two maps Both land cover maps have been rasterised with a resolution of 50 m for the present analysis. These 50m 50m cells are called pixels. A visual inspection of the overlay of both land cover maps in UTM 32 indicates that the co-location accuracy is generally within 50 m (Figure 2), and nearly always within 100 m, i.e. within CORINE Land Cover specifications. Figure 2: CORINE Land Cover poygons with ILC map rasterised at 50 m. If we disregard co-location inaccuracy, we can describe the disagreement through a matrix (often named confusion matrix) with elements A = area of the pixels with code c in CLC and c in ILC. This confusion matrix would be a naïf description of the disagreement. To eliminate the part of the disagreement due to co-location tolerance (100 m), we eliminate a 2 pixels wide on each side of the CORINE Land Cover polygon borders. This corresponds to eliminating a 200m wide corridor around the polygon limits (Figure 3). cc' CORINE Land Cover Figure 3: Buffer of 2 pixels eliminated around CORINE Land Cover limits with rasterised CORINE Land Cover ILC
3.1 Pixelwise (naïf) commission and omission disagreement. The accuracy of a rasterised land cover map can be assessed through the confusion matrix, that gives the number of pixels A for which the true land cover is c and the land cover cc' given by the map is c. The omission error for class c refers to the proportion of pixels for which the truth is c and the map gives a different class. The commission error is the proportion of pixels represented as c by the map, but having a different true land cover. Here the usual terms confusion matrix and commission and omission errors are substituted by disagreement matrix and commission and omission disagreement to stress the fact that two different land cover maps are compared: none of the maps can be considered as the truth and part of the disagreement can be explained by the different scales; in this case it should not be considered as an error. The agreement between two land cover maps can be computed ast he overall % of pixels with coinciding codes, i.e. in the main diagonal of the disagreement matrix. The kappa statistic (Bishop et al, 1975) is a better measure of agreement: its value is close to 0 when the coincidences are random and 1 when both maps coincide perfectly. Table 3 shows both agreement measures before and after removing s. The comparison gives an idea of the part of the disagreement that can be attributed to co-location inaccuracy. Total area % coincident kappa *1000 ha no 199 56 0.42 200 m 112 67 0.52 Table 3: Agreement from a confusion matrix between CORINE and ILC. Commission Omission No 200m No 200m Urban dense 89 87 46 37 Urban discont. 59 62 60 57 Indus-comm 64 62 36 21 Road-rail 96 99 90 88 Mines 40 32 17 3 Arable non irrigated 51 39 24 19 Vineyards 73 76 54 43 Fruits 46 47 30 22 Olive 49 41 35 24 Pastures 76 78 92 92 Arable+perm 80 81 94 94 Complex agr 59 55 87 87 Agric + nat.veg 78 75 92 92 Forest broad leaved 27 16 14 10 Coniferous 60 58 33 18 Forest mixed 63 64 87 85 Nat. Grass 86 88 60 67 Moors-heath 69 70 63 55 Wood-shrub 87 85 96 95 Water bodies 45 63 30 0 Table 4: Commission and omission disagreement by pixel (in %).
Table 4 reports commission and omission disagreements with and without for categories that have at least 100 ha for both maps. We should stress again that pixelwise disagreement can be very misleading: both maps can be perfectly consistent with each other, and have a high % of disagreement by pixel because they represent the same reality at different scale. We shall see below that both maps have a very good agreement when the scale effect is removed. 3.2 Recoding into 9 classes. To simplify the analysis of different types of disagreement, we have recoded both maps into 9 major classes. Figure 1 shows that both land cover maps have a similar pattern, with the obvious smoothing effect for CORINE Land Cover. Tables 5, 6 and 7 below report main agreement parameters when a pixelwise overlay is performed. Total area % coincident kappa *1000 ha no 199 68 0.51 200 m 125 77 0.61 Table 5: Agreement between CORINE and ILC with 9 classes. Commis sion Omissio n No 200m No 200m Urban 47.4 44.4 22.3 8.2 Arable 50.4 40.2 20.1 14.7 Permanent crops 51.6 49.9 33.5 23.6 Pastures 76.4 79.5 92.3 92.3 Heterogeneous 43.5 35.8 80.8 80.4 Forest 17.3 9.2 10.0 6.0 Other nat. veg. 61.5 63.6 69.5 66.3 Open spaces 80.2 93.8 42.2 18.4 Marsh and water 63.9 75.4 35.0 32.0 Table 6: Commission and omission disagreement with 9 classes (%) CORINE ISTAT Land Cover ha Urban Arable Perm. Pastures Heteroge Forest Other Open Water Total crops neous nat.veg. spaces Urban 2601 105 10 0 46 34 29 0 10 2834 Arable 532 13355 292 89 513 545 200 5 120 15650 Perm. crops 126 258 3160 18 219 299 53 0 4 4136 Pastures 66 1597 25 201 277 179 273 5 2 2623 Heterogeneous 1043 6442 2223 406 3162 2188 660 7 9 16141 Forest 180 393 527 163 435 72071 2916 16 8 76710 Other nat.veg. 129 178 68 104 262 3960 2405 25 3 7133 Open spaces 0 0 0 0 0 1 0 4 0 5 Water 0 1 0 0 0 0 17 6 51 74 Total 4677 22329 6305 981 4914 79277 6553 68 207 125306 Table 7: Pixelwise disagreeement matrix with 9 classes and 200 m.
Off-diagonal figures correspond at first sight to disagreements, but they do not take into account the scale effect. For example the 2 ha in the case pastures water correspond to a small pond in a pasture area, and CORINE Land Cover respects its specifications not reporting it; therefore it cannot be considered as a real disagreement. There are hundreds of other false disagreements in the test site that distort Table 7. 3.3 Scale-corrected thematic disagreement by CORINE Land Cover polygon. If a CLC polygon has 80% of arable land, it is correctly classified according to the specifications of this land cover map. Similar criteria apply for the main CLC categories, excepting the ones that correspond to heterogeneous landscapes. We have followed the procedure described below to remove the part of the disagreement between both land cover maps due to the different scales: The assessment is made on the nomenclature grouped into 9 classes. The class heterogeneous agriculture does not follow the rules below. We eliminate as well the class burnt areas ; A of 200 m is removed around CORINE Land Cover polygon borders (100 at each side) to avoid counting location inaccuracy as disagreement; If more than 70% of the pixels have the same code for the ILC map, the whole polygon is considered in agreement; If the % of different code is between 30 and 70, the polygon is considered partially in disagreement; If less than 30% of the pixels have a different code for the ILC map, the whole polygon is considered in disagreement; For example if a polygon of 100 ha (after removing the ) has been coded arable in CORINE Land Cover and ILC reports: 85 ha arable and 15 permanent crops, it will no contribute to the disagreement. 60 ha arable and 40 permanent crops, it will contribute as 40 ha disagreement 20 ha arable and 80 permanent crops, it will contribute as 100 ha disagreement CORINE Land Cover area excluding (Kha) % disagreement Urban 2.8 2.2 Arable 15.2 1.4 Permanent crops 3.9 6.4 Pastures 2.3 93.6 Heterogeneous 13.0. Forest 76.2 0.3 Other nat. veg. 6.8 68.6 Open spaces 0.2 0 Marsh and water 0.1 32 Table 8: Thematic disagreement with a grouped nomenclature. Table 8 reports the disagreement rates. This table confirms several expected facts: The agreement is very good for urban, arable, forest, and open spaces.
There are major discrepancies on pasture and natural vegetation, that represent 7.5% of the area (excluding s). This divergence probably comes from the difficulty to interpret the nomenclature. Other figures are more surprising but may have some explanation: The disagreement for marsh and water comes from a single marsh polygon. The low disagreement of permanent crops increases our confidence in CLC for this land cover class, especially difficult to photo-interpret. References Bishop Y., Fienberg S., Holland P., 1975, Discrete Multivariate Analysis, M.I.T. press, Cambridge, Ma. EC, 1993, CORINE Land Cover; guide technique, Report EUR 12585EN. Office for Publications of the European Communities. Luxembourg,. 144 pp Gallego, F.J. Carfagna, E., Peedell S., 1999, The use of CORINE Land Cover to improve area frame survey estimates. Research in Official Statistics, Vol 2, no 2, pp. 99-122 ISTAT, 1998, - Capitolato tecnico per il progetto pilota del database sull'uso e la copertura del suolo in scala 1:25.000 su un area test. Rome Napolitano P., Carbonetti G., Gallego J., 2000, Accuracy Assessment and Validation of a Land Use-Land Cover Database in Arezzo (Italy). Proceedings of Accuracy 2000, Amsterdam, July 12-14, 2000, pp. 233-236.